Documentation Fetcher & RAG Search

A modular system for fetching API documentation and enabling semantic search via RAG (Retrieval-Augmented Generation). Designed to give AI coding assistants like Claude access to up-to-date documentation from any project.

Features

• Fetch Documentation: Download complete documentation from API providers in markdown format
• Semantic Search: Hybrid search combining vector embeddings with keyword matching
• MCP Server: Expose search as tools accessible from Claude Code in any project
• Modular Design: Easy to add new documentation sources

Supported Documentation Sources

Source	Documents	Description
Gemini	~2000	Google Gemini API - LLM, function calling, embeddings, multimodal
FastMCP	~1900	FastMCP framework - MCP servers, tools, resources, authentication

Quick Start

Prerequisites

• Python 3.12+
• Ollama with bge-m3 model
• Claude Code (for MCP integration)

Installation

# Clone the repository
git clone <repository-url>
cd documentation

# Create virtual environment
python3.12 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Pull the embedding model
ollama pull bge-m3

Fetch & Index Documentation

# Fetch documentation
python -m src.main fetch gemini
python -m src.main fetch fastmcp

# Index for search (requires Ollama running)
python -m src.rag.index gemini
python -m src.rag.index fastmcp

Search Documentation

# Search Gemini docs
python -m src.main search "function calling"

# Search FastMCP docs
python -m src.main search "how to create a tool" -c fastmcp

# More results
python -m src.main search "rate limits" -n 10

MCP Server Integration

The MCP server exposes documentation search as tools that Claude Code can use from any project.

Install in Claude Code

IMPORTANT: MCP configuration requires absolute paths. The cwd field is NOT supported by Claude Code.

Option 1: Using Claude CLI (recommended)

# Replace /path/to/documentation with your actual absolute path
claude mcp add docs-search --scope user --transport stdio -- \
  /path/to/documentation/.venv/bin/python \
  /path/to/documentation/src/mcp_server.py

Option 2: Add to ~/.claude.json manually

{
  "mcpServers": {
    "docs-search": {
      "command": "/path/to/documentation/.venv/bin/python",
      "args": ["/path/to/documentation/src/mcp_server.py"]
    }
  }
}

Common mistakes to avoid:

• Do NOT use cwd - it's not a valid MCP configuration field
• Do NOT use relative paths - they resolve from the caller's directory
• Do NOT use -m src.mcp_server - this requires being in the project directory

Verify Installation

# Check server is registered
claude mcp list

# In Claude Code, check connection status
/mcp

Available Tools

Tool	Description
search_docs(query, collection, num_results)	Search documentation with hybrid semantic + keyword search
list_collections()	List available documentation collections

Available Resources

Resource URI	Description
docs://collections	JSON list of all collections
docs://gemini/pages	List of all Gemini documentation pages
docs://fastmcp/pages	List of all FastMCP documentation pages
docs://gemini/search-help	Search tips for Gemini docs
docs://fastmcp/search-help	Search tips for FastMCP docs

Usage from Claude Code

Once installed, you can ask Claude from any project:

• "Search the gemini docs for function calling"
• "What documentation collections are available?"
• "Search fastmcp for how to create tools"
• "Find rate limit information in gemini docs"

Project Structure

documentation/
├── src/
│   ├── main.py                 # CLI entry point
│   ├── mcp_server.py           # MCP server for Claude Code
│   ├── core/
│   │   ├── fetcher.py          # HTTP/markdown fetching
│   │   └── parser.py           # Navigation parsing
│   ├── modules/
│   │   ├── base.py             # Abstract base class
│   │   ├── gemini/             # Gemini documentation module
│   │   └── fastmcp/            # FastMCP documentation module
│   └── rag/
│       ├── chunker.py          # Markdown-aware chunking
│       ├── embedder.py         # Ollama bge-m3 embeddings
│       ├── sqlite_store.py     # SQLite + sqlite-vec vector store
│       ├── search.py           # Hybrid search with RRF
│       ├── query_expander.py   # Multi-query expansion (LLM)
│       ├── reranker.py         # Cross-encoder reranking
│       └── index.py            # Indexing CLI
├── output/                     # Fetched documentation
│   ├── gemini/
│   └── fastmcp/
├── data/
│   └── docs.db                 # SQLite vector database
├── requirements.txt
└── README.md

Adding New Documentation Sources

1. Create a new module in src/modules/<name>/:

# src/modules/example/config.py
BASE_URL = "https://docs.example.com"
SITEMAP_URL = "https://docs.example.com/sitemap.xml"
MARKDOWN_SUFFIX = ".md"  # or ".md.txt" for Google sites

# src/modules/example/module.py
from src.modules.base import BaseModule

class ExampleModule(BaseModule):
    @property
    def name(self) -> str:
        return "example"

    def get_doc_urls(self) -> list[NavLink]:
        # Parse sitemap or navigation
        ...

    def fetch_page(self, url: str) -> str:
        # Fetch markdown content
        ...

2. Register in src/main.py:

from src.modules.example.module import ExampleModule

# In fetch_command():
elif args.module == "example":
    module = ExampleModule()
    module.run(output_dir)

3. Add to KNOWN_COLLECTIONS in src/mcp_server.py

4. Fetch and index:

python -m src.main fetch example
python -m src.rag.index example

How It Works

Fetching

1. Parse navigation/sitemap to discover documentation pages
2. Fetch each page in markdown format (using source-specific tricks like .md.txt suffix)
3. Save with source URL metadata

Indexing

1. Chunk markdown by headers (preserving code blocks)
2. Generate embeddings via Ollama bge-m3 (1024 dimensions)
3. Store in SQLite with sqlite-vec (vectors) and FTS5 (keywords)

Searching

1. Generate query embedding
2. Perform semantic search (sqlite-vec vector similarity)
3. Perform keyword search (FTS5 BM25)
4. Combine with Reciprocal Rank Fusion (RRF)
5. Optionally expand query with LLM variations
6. Optionally rerank with cross-encoder
7. Return ranked results with source URLs

Configuration

Environment Variables

Variable	Description	Default
OLLAMA_HOST	Ollama server URL	http://localhost:11434

SQLite Database

Vector database stored in data/docs.db. Each documentation source gets its own collection within the database.

Development

# Run tests
python -m pytest

# Check MCP server
claude mcp list

# Test search functionality
python -m src.rag.search

Troubleshooting

"Ollama connection failed"

# Make sure Ollama is running
ollama serve

# Pull the embedding model
ollama pull bge-m3

"No results found"

# Check if collection is indexed
python -m src.rag.index --status gemini

# Re-index if needed
python -m src.rag.index --clear gemini

MCP server not connecting

# Check server status
claude mcp list

# Reinstall
claude mcp remove docs-search
fastmcp install claude-code src/mcp_server.py --name docs-search

License

MIT

Credits

• Ollama - Local LLM and embeddings
• sqlite-vec - Vector search for SQLite
• FastMCP - MCP server framework